A new Nature Medicine study suggests the biggest obstacle to clinical AI in emergency departments is keeping doctors engaged, not the accuracy of the models.
The study put SHAKED, a decision-support system assembled from several large language models, through a DECIDE-AI stage 1 evaluation at a tertiary emergency department. A month-long run tracked 1,138 patients across twin units, one staffed with SHAKED and the other sticking to standard rotations.
Clinical adoption of the system fell from 68% to 30% over the study period. Workload-sensitive disengagement drove the decline, with each additional shift hour cutting the odds of use by roughly a quarter. Physicians did gravitate toward SHAKED for radiology consultations, where the odds of use were about three times higher.
The safety picture was clean: reviewers flagged no adverse events, and clinicians judged 99 of 100 sampled outputs appropriate for use. Emergency department length of stay did not differ between the two wings, both averaging 4.9 hours, and there was a non-significant trend toward shorter consultation cycle times.
The takeaway from the team: keeping clinicians engaged, not model precision, may be what decides whether clinical AI succeeds in emergency medicine. They stop short of endorsing deployment, saying the results inform randomized trial design. The study is registered as NCT06902675.
